Including SAR assessing metrics in despeckling networks

Abstract

In the recent years many deep learning (DL) based solutions for SAR image despeckling have been proposed. These solutions vary from different aspects: training approaches, architectures and cost functions. Even if different cost functions have been proposed, from simple to multi-objective ones, the training is still mainly based on the use of euclidean norms as loss term. In this work the inclusion of SAR theoretical background in the cost function is exploited. In particular, assessing metrics specifically designed for the evaluation of despeckling filters are considered as cost function for the training of DL solutions. Results on validation dataset and on real data motivate to further investigate in this direction.